for written notes on this lecture, please read chapter 14...

For written notes on this lecture, please read chapter 14 of The Practical Bioinformatician.

CS2220: Introduction to Computational Biology

Lecture 4: Gene Expression Analysis

Limsoon Wong

Copyright 2012 © Limsoon Wong

2

Plan

• Microarray background

• Gene expression profile classification

• Gene expression profile clustering

• Normalization

• Extreme sample selection

• Intersection Analysis

Background on Microarrays


4

What is a Microarray?

• Contain large number of DNA molecules spotted

on glass slides, nylon membranes, or silicon

wafers

• Detect what genes are being expressed or found

in a cell of a tissue sample

• Measure expression of thousands of genes

simultaneously


5

Affymetrix GeneChip Array

http://www.dnalc.org/ddnalc/resources/dnachip.html


6

quartz is washed to ensure uniform

hydroxylation across its surface and to

attach linker molecules

exposed linkers become deprotected and

are available for nucleotide coupling

Making Affymetrix GeneChip Array

Exercise: What is the other commonly used

type of microarray? How is that one different

from Affymetrix’s?


7

Gene Expression Measurement

by Affymetrix GeneChip Array

Click to

watch an

interesting

movie

explaining the

working of

microarray

http://www.bio.davidson.edu/Courses/genomics/chip/chip.html


8

A Sample Affymetrix GeneChip

Data File (U95A)


9

Some Advice on

Affymetrix Gene Chip Data

• Ignore AFFX genes

– These genes are control genes

• Ignore genes with “Abs Call” equal to “A” or “M”

– Measurement quality is suspect

• Upperbound 40000, lowerbound 100

– Accuracy of laser scanner

• Deal with missing values Exercise: Suggest 2 ways

to deal with missing value


10

Type of Gene Expression Datasets

Class Gene1 Gene2 Gene3 Gene4 Gene5 Gene6 Gene7 .....

Sample1 Cancer 0.12 -1.3 1.7 1.0 -3.2 0.78 -0.12

Sample2 Cancer 1.3

.

~Cancer

SampleN ~Cancer

1000 - 100,000 columns

100-500 rows

Gene-Conditions or Gene-Sample (numeric or discretized)

Gene-Sample-Time Gene-Time

time

expre

ssio

n le

ve

l


11


1000 - 100,000 columns

100-500 rows


Class Gene1 Gene2 Gene3 Gene4 Gene5 Gene6 Gene7 .....

Sample1 Cancer 1 0 1 1 1 0 0

Sample2 Cancer 1

.

~Cancer

SampleN ~Cancer


time

expre

ssio

n le

ve

l


12

Application: Disease Subtype Diagnosis

???

malign

malign

malign

malign

benign

benign

benign

benign

??? ???

genes

sam

ple

s


13

Application: Treatment Prognosis

???

NR

NR

NR

NR

R

R

R

R

??? ???

genes

sam

ple

s


14


1000 - 100,000 columns

100-500 rows


Gene1 Gene2 Gene3 Gene 4 Gene5 Gene6 Gene7

Cond1 0.12 -1.3 1.7 1.0 -3.2 0.78 -0.12

Cond2 1.3

.

CondN


time

expre

ssio

n le

ve

l


15

Application: Drug Action Detection

Normal

Normal

Normal

Normal

Drug

Drug

Drug

Drug

genes

condit

ions

Which group of genes are the drug affecting on?

Gene Expression Profile Classification

Diagnosis of Childhood Acute

Lymphoblastic Leukemia and Optimization

of Risk-Benefit Ratio of Therapy


17

• The subtypes look similar

• Conventional diagnosis

– Immunophenotyping

– Cytogenetics

– Molecular diagnostics

• Unavailable in most

ASEAN countries

Childhood ALL

• Major subtypes: T-ALL,

E2A-PBX, TEL-AML, BCR-

ABL, MLL genome

rearrangements,

Hyperdiploid>50

• Diff subtypes respond

differently to same Tx

• Over-intensive Tx

– Development of

secondary cancers

– Reduction of IQ

• Under-intensiveTx

– Relapse


18

Mission

• Conventional risk assignment procedure requires

difficult expensive tests and collective judgement

of multiple specialists

• Generally available only in major advanced

hospitals

Can we have a single-test easy-to-use platform

instead?


19

Single-Test Platform of

Microarray & Machine Learning


20

Overall Strategy

• For each subtype, select

genes to develop

classification model for

diagnosing that subtype

• For each subtype, select

genes to develop

prediction model for

prognosis of that subtype

Diagnosis

of subtype

Subtype-

dependent

prognosis

Risk-

stratified

treatment

intensity


21

Subtype Diagnosis by PCL

• Gene expression data collection

• Gene selection by 2

• Classifier training by emerging pattern

• Classifier tuning (optional for some machine

learning methods)

• Apply classifier for diagnosis of future cases by

PCL


22

Childhood ALL Subtype

Diagnosis Workflow

A tree-structured

diagnostic

workflow was

recommended by

our doctor

collaborator


23

Training and Testing Sets


24

Signal Selection Basic Idea

• Choose a signal w/ low intra-class distance

• Choose a signal w/ high inter-class distance


25

Signal Selection by 2


26

Emerging Patterns

• An emerging pattern is a set of conditions

– usually involving several features

– that most members of a class satisfy

– but none or few of the other class satisfy

• A jumping emerging pattern is an emerging

pattern that

– some members of a class satisfy

– but no members of the other class satisfy

• We use only jumping emerging patterns


27

Examples

Reference number 9: the expression of gene 37720_at > 215

Reference number 36: the expression of gene 38028_at 12

Patterns Frequency (P) Frequency(N)

{9, 36} 38 instances 0

{9, 23} 38 0

{4, 9} 38 0

{9, 14} 38 0

{6, 9} 38 0

{7, 21} 0 36

{7, 11} 0 35

{7, 43} 0 35

{7, 39} 0 34

{24, 29} 0 34

Easy interpretation


28

PCL: Prediction by Collective Likelihood


29

PCL Learning

Top-Ranked EPs in

Positive class

Top-Ranked EPs in

Negative class

EP1P (90%)

EP2P (86%)

.

.

EPnP (68%)

EP1N (100%)

EP2N (95%)

.

.

EPnN (80%)

The idea of summarizing multiple top-ranked EPs is intended

to avoid some rare tie cases


30

PCL Testing

ScoreP = EP1P’ / EP1

P + … + EPkP’ / EPk

P

Most freq EP of pos class

in the test sample

Most freq EP of pos class

Similarly,

ScoreN = EP1N’ / EP1

N + … + EPkN’ / EPk

N

If ScoreP > ScoreN, then positive class,

Otherwise negative class


31

Accuracy of PCL (vs. other classifiers)

The classifiers are all applied to the 20 genes selected

by 2 at each level of the tree


32

Understandability of PCL

• E.g., for T-ALL vs. OTHERS, one ideally

discriminatory gene 38319_at was found,

inducing these 2 EPs

• These give us the diagnostic rule


33

Multidimensional Scaling Plot

for Subtype Diagnosis

Obtained by performing PCA on the 20 genes chosen for each level


34

Childhood ALL Cure Rates

• Conventional risk

assignment procedure

requires difficult

expensive tests and

collective judgement of

multiple specialists

Not available in less

advanced ASEAN

countries


35

Childhood ALL Treatment Cost

• Treatment for childhood ALL over 2 yrs

– Intermediate intensity: US$60k

– Low intensity: US$36k

– High intensity: US$72k

• Treatment for relapse: US$150k

• Cost for side-effects: Unquantified


36

Current Situation

(2000 new cases/yr in ASEAN)

• Intermediate intensity

conventionally applied in

less advanced ASEAN

countries

• Over intensive for 50% of

patients, thus more side

effects

• Under intensive for 10% of

patients, thus more relapse

• US$120m (US$60k * 2000)

for intermediate intensity tx

• US$30m (US$150k * 2000 *

10%) for relapse tx

• Total US$150m/yr plus un-

quantified costs for dealing

with side effects


37

Using Our Platform

• Low intensity applied to

50% of patients

• Intermediate intensity to

40% of patients

• High intensity to 10% of

patients

Reduced side effects

Reduced relapse

75-80% cure rates

• US$36m (US$36k * 2000 *

50%) for low intensity

• US$48m (US$60k * 2000 *

40%) for intermediate

intensity

• US$14.4m (US$72k * 2000 *

10%) for high intensity

• Total US$98.4m/yr

Save US$51.6m/yr


38

A Nice Ending…

• Asian Innovation

Gold Award 2003

Gene Expression Profile Clustering

Novel Disease Subtype Discovery


40

Is there a new subtype?

• Hierarchical

clustering of

gene expression

profiles reveals a

novel subtype of

childhood ALL

Exercise: Name and describe

one bi-clustering method


41

Hierarchical Clustering

• Assign each item to its own cluster

– If there are N items initially, we get N clusters,

each containing just one item

• Find the “most similar” pair of clusters, merge

them into a single cluster, so we now have one

less cluster

– ―Similarity‖ is often defined using

• Single linkage

• Complete linkage

• Average linkage

• Repeat previous step until all items are clustered

into a single cluster of size N


42

Single, Complete, & Average Linkage

Single linkage defines distance

betw two clusters as min distance

betw them

Complete linkage defines distance

betw two clusters as max distance betw

them

Exercise: Give definition of “average linkage”

Image source: UCL Microcore Website


43

Some Patient Samples

• Does Mr. A have cancer?

malign

malign

malign

malign

benign

benign

benign

benign

genes

sam

ple

s

??? Mr. A:


44

Let’s rearrange the rows…

• Does Mr. A have cancer?

genes

sam

ple

s

malign

malign

malign

malign

benign

benign

benign

benign

??? Mr. A:


45

and the columns too…

malign

malign

malign

malign

benign

benign

benign

benign

genes

sam

ple

s

??? Mr. A:

Normalization


47

Sometimes, a gene expression study

may involve batches of data collected

over a long period of time…

0

10

20

30

40

50

60

70

Jan-0

4

Apr-

04

Jul-04

Oct-

04

Jan-0

5

Apr-

05

Jul-05

Oct-

05

Jan-0

6

Apr-

06

Jul-06

Oct-

06

Jan-0

7

Apr-

07

Jul-07

Oct-

07

Jan-0

8

Apr-

08

Jul-08

Oct-

08

Jan-0

9

Apr-

09

Jul-09

Oct-

09

Jan-1

0

Time Span of Gene Expression Profiles

Image credit: Dong Difeng


48

In such a case, batch effect may be

severe… to the extent that you can

predict the batch that each sample

comes!

Need normalization to correct for batch effect

Image credit: Dong Difeng


49

Approaches to Normalization

• Aim of

normalization:

Reduce variance

w/o increasing bias

• Scaling method

– Intensities are scaled

so that each array

has same ave value

– E.g., Affymetrix’s

• Xform data so that

distribution of

probe intensities is

same on all arrays

– E.g., (x ) /

• Quantile

normalization


50

Quantite Normalization

• Given n arrays of length p,

form X of size p × n where

each array is a column

• Sort each column of X to

give Xsort

• Take means across rows

of Xsort and assign this

mean to each elem in the

row to get X’sort

• Get Xnormalized by arranging

each column of X’sort to

have same ordering as X

• Implemented in some

microarray s/w, e.g.,

EXPANDER


51

After quantile

normalization

Selection of Patient Samples and Genes

for Disease Prognosis


53

Gene Expression Profile

+ Clinical Data

Outcome Prediction

• Univariate & multivariate Cox survival analysis (Beer et al 2002, Rosenwald et al 2002)

• Fuzzy neural network (Ando et al 2002)

• Partial least squares regression (Park et al 2002)

• Weighted voting algorithm (Shipp et al 2002)

• Gene index and “reference gene” (LeBlanc et al 2003)

• ……


54

Our Approach

“extreme”

sample

selection

ERCOF


55

Short-term Survivors v.s. Long-term Survivors

T: sample

F(T): follow-up time

E(T): status (1:unfavorable; 0: favorable)

c1 and c2: thresholds of survival time

Short-term survivors

who died within a short

period

F(T) < c1 and E(T) = 1

Long-term survivors

who were alive after a

long follow-up time

F(T) > c2

Extreme Sample Selection


56

ERCOF Entropy-

Based Rank

Sum Test &

Correlation

Filtering

Remove genes with

expression values w/o

cut point found (can’t

be discretized)

Calculate Wilcoxon rank

sum w(x) for gene x.

Remove gene x if w(x)

[clower, cupper]

Group features by

Pearson Correlation

For each group, retain

the top 50% wrt class

entropy


57

Linear Kernel SVM regression function

bixTKyaTG i

i

i ))(,()(

T: test sample, x(i): support vector,

yi: class label (1: short-term survivors; -1: long-term survivors)

Transformation function (posterior probability)

)(1

1)(

TGe

TS

))1,0()(( TS

S(T): risk score of sample T

Risk Score Construction


58

Diffuse Large B-Cell Lymphoma

• DLBC lymphoma is the

most common type of

lymphoma in adults

• Can be cured by

anthracycline-based

chemotherapy in 35 to 40

percent of patients

DLBC lymphoma

comprises several

diseases that differ in

responsiveness to

chemotherapy

• Intl Prognostic Index (IPI)

– age, ―Eastern Cooperative

Oncology Group‖ Performance

status, tumor stage, lactate

dehydrogenase level, sites of

extranodal disease, ...

• Not very good for stratifying

DLBC lymphoma patients for

therapeutic trials

Use gene-expression

profiles to predict outcome

of chemotherapy?


59

Rosenwald et al., NEJM 2002

• 240 data samples

– 160 in preliminary group

– 80 in validation group

– each sample described by 7399 microarray

features

• Rosenwald et al.’s approach

– identify gene: Cox proportional-hazards model

– cluster identified genes into four gene signatures

– calculate for each sample an outcome-predictor

score

– divide patients into quartiles according to score


60

Knowledge Discovery from Gene

Expression of ―Extreme‖ Samples

“extreme”

sample

selection:

< 1 yr vs > 8 yrs

knowledge

discovery

from gene

expression

240

samples

80

samples 26 long-

term survivors

47 short-

term survivors

7399

genes

84

genes

T is long-term if S(T) < 0.3

T is short-term if S(T) > 0.7


61

73 25 47+1(*) Informative

160 72 88 Original DLBCL

Alive Dead

Total Status Data set Application

Number of samples in original data and selected informative training set.

(*): Number of samples whose corresponding patient was dead at the end

of follow-up time, but selected as a long-term survivor.

Discussions: Sample Selection


62

84(1.7%) Phase II

132(2.7%) Phase I

4937(*) Original

DLBCL Gene selection

Number of genes left after feature filtering for each phase.

(*): number of genes after removing those genes who were

absent in more than 10% of the experiments.

Discussions: Gene Identification


63

p-value of log-rank test: < 0.0001

Risk score thresholds: 0.7, 0.3

Kaplan-Meier Plot for 80 Test Cases


64

(A) IPI low,

p-value = 0.0063

(B) IPI intermediate,

p-value = 0.0003

Improvement Over IPI


65

(A) W/o sample selection (p =0.38) (B) With sample selection (p=0.009)

No clear difference on the overall survival of the 80 samples in the validation

group of DLBCL study, if no training sample selection conducted

Merit of ―Extreme‖ Samples


66

About the Inventor: Huiqing Liu

• Huiqing Liu

– PhD, NUS, 2004

– Currently Senior

Scientist at Centocor

– Asian Innovation

Gold Award 2003

– New Jersey Cancer

Research Award for

Scientific Excellence

2008

– Gallo Prize 2008

Beyond Disease Diagnosis & Prognosis


68

Beyond Classification of

Gene Expression Profiles

• After identifying the candidate genes by feature

selection, do we know which ones are causal

genes, which ones are surrogates, and which are

noise? Diagnostic ALL BM samples (n=327)

3 -3 -2 -1 0 1 2 = std deviation from mean

Ge

nes

fo

r c

las

s

dis

tin

cti

on

(n

=2

71

)

TEL-AML1 BCR-

ABL

Hyperdiploid >50 E2A-

PBX1

MLL T-ALL Novel


69

Gene Regulatory Circuits

• Genes are “connected”

in “circuit” or network

• Expr of a gene in a

network depends on

expr of some other

genes in the network

• Can we “reconstruct”

the gene network from

gene expression and

other data? Source: Miltenyi Biotec


70

• Each disease subtype has underlying cause

There is a unifying biological theme for genes

that are truly associated with a disease subtype.

• Uncertainty in reliability of selected genes can be

reduced by considering molecular functions and

biological processes associated with the genes

• The unifying biological theme is basis for

inferring the underlying cause of disease subtype

Hints to extend reach of prediction


71

Intersection Analysis

• Intersect the list of

differentially expressed

genes with a list of genes

on a pathway

• If intersection is

significant, the pathway is

postulated as basis of

disease subtype or

treatment response

Caution:

• Initial list of differentially

expressed genes is

defined using test

statistics with arbitrary

thresholds

• Diff test statistics and diff

thresholds result in a diff

list of differentially

expressed genes

Outcome may be unstable Exercise: What is a good test

statistics to determine if the

intersection is significant?

Gene Interaction Prediction


73

Beyond Classification of

Gene Expression Profiles

• After identifying the candidate genes by feature

selection, do we know which ones are causal

genes and which ones are surrogates?

Diagnostic ALL BM samples (n=327)

3 -3 -2 -1 0 1 2 = std deviation from mean

Ge

nes

fo

r c

las

s

dis

tin

cti

on

(n

=2

71

)

TEL-AML1 BCR-

ABL

Hyperdiploid >50 E2A-

PBX1

MLL T-ALL Novel


74

Gene Regulatory Circuits

• Genes are “connected” in

“circuit” or network

• Expression of a gene in a

network depends on

expression of some other

genes in the network

• Can we reconstruct the

gene network from gene

expression data?


75

Key Questions

• For each gene in the network:

• Which genes affect it?

• How they affect it?

– Positively?

– Negatively?

– More complicated ways?


76

Some Techniques

• Bayesian Networks

– Friedman et al., JCB 7:601--620, 2000

• Boolean Networks

– Akutsu et al., PSB 2000, pages 293--304

• Differential equations

– Chen et al., PSB 1999, pages 29--40

• Classification-based method

– Soinov et al., ―Towards reconstruction of gene

network from expression data by supervised

learning‖, Genome Biology 4:R6.1--9, 2003


77

A Classification-Based Technique Soinov et al., Genome Biology 4:R6.1-9, Jan 2003

• Given a gene expression matrix X

– each row is a gene

– each column is a sample

– each element xij is expression of gene i in sample j

• Find the average value ai of each gene i

• Denote sij as state of gene i in sample j,

– sij = up if xij > ai

– sij = down if xij ai


78

A Classification-Based Technique Soinov et al., Genome Biology 4:R6.1-9, Jan 2003

• To see whether the state of

gene g is determined by

the state of other genes

– see whether sij | i g

can predict sgj

– if can predict with high

accuracy, then ―yes‖

– Any classifier can be

used, such as C4.5, PCL,

SVM, etc.

• To see how the state of

gene g is determined by

the state of other genes

– apply C4.5 (or PCL or

other ―rule-based‖

classifiers) to predict sgj

from sij | i g

– and extract the decision

tree or rules used


79

Advantages of this method

• Can identify genes affecting a target gene

• Don’t need discretization thresholds

• Each data sample is treated as an example

• Explicit rules can be extracted from the classifier

(assuming C4.5 or PCL)

• Generalizable to time series

Concluding Remarks


81

Bcr-Abl

• Targeted drug dev

– Know what

molecular effect you

want to achieve

• E.g., inhibit a

mutated form of a

protein

– Engineer a

compound that

directly binds and

causes the desired

effect

• Gleevec (imatinib)

– 1st success for real drug

– Targets Bcr-Abl fusion

protein (ie, Philadelphia

chromosome, Ph)

– NCI summary of clinical

trial of imatinib for ALL

at

http://www.cancer.gov/clini

caltrials/results/ALLimati

nib1109/print

http://www.cancer.gov/clinicaltrials/results/ALLimatinib1109/print




82

What have we learned?

• Technologies

– Microarray

– PCL, ERCOF

• Microarray applications

– Disease diagnosis by supervised learning

– Subtype discovery by unsupervised learning

• Important tactics

– Extreme sample selection

– Intersection analysis, Gene network reconstruction


83

Useful Gene Expression

Analysis Packages

• EXPANDER (EXPression Analyser & DisplayER)

– Ron Shamir @ Tel Aviv University

– http://acgt.cs.tau.ac.il/expander

• BRB-Array Tools

– Biometric Research Branch @ National Cancer

Center, USA

– http://linus.nci.nih.gov/BRB-ArrayTools.html

http://acgt.cs.tau.ac.il/expander

http://linus.nci.nih.gov/BRB-ArrayTools.html



Any Question?


85

References

• E.-J. Yeoh et al., ―Classification, subtype discovery, and

prediction of outcome in pediatric acute lymphoblastic leukemia

by gene expression profiling‖, Cancer Cell, 1:133--143, 2002

• H. Liu, J. Li, L. Wong. Use of Extreme Patient Samples for

Outcome Prediction from Gene Expression Data. Bioinformatics,

21(16):3377--3384, 2005.

• L.D. Miller et al., ―Optimal gene expression analysis by

microarrays‖, Cancer Cell 2:353--361, 2002

• J. Li, L. Wong, ―Techniques for Analysis of Gene Expression‖,

The Practical Bioinformatician, Chapter 14, pages 319—346,

WSPC, 2004

• B. Bolstad et al. ―A comparison of normalization methods for

high density oligonucleotide array data based on variance and

bias‖. Bioinformatics, 19:185–193. 2003

psZ/jim-allen2002-final.pdf

for written notes on this lecture, please read chapter 14...

Documents